pi0.5-DROID fine-tuned on digital-twin rollouts (planner data vs steering data)
LoRA fine-tunes of openpi's pi05_droid, trained only on simulated rollouts from MuJoCo twins of DROID scenes and of
a real Franka bench. Research checkpoints from an in-progress comparison; not validated on real hardware.
Each folder holds params/ and assets/droid/ (the unchanged DROID norm stats).
| folder | data | scenes |
|---|---|---|
ep4_steer_s4000 |
steering (lookahead over unmodified pi0.5 samples), 52 episodes | one DROID-episode twin |
all_setups_planner_s<step> |
scripted planner, 206 episodes | six twins including the bench twin |
all_setups_steer_s<step> |
steering, 206 episodes | same six twins |
<step> is the training step of the checkpoint.
Use
The model must be built with the LoRA variants (paligemma_variant="gemma_2b_lora",
action_expert_variant="gemma_300m_lora") on top of the pi05_droid config. A small server that does this and
speaks openpi's websocket protocol is scripts/svp_serve.py in https://github.com/Sumo048/scan_sim (branch
steer-vs-planner); REAL_ROBOT_AGENT.md there describes the tasks, instruction strings and evaluation protocol.
Interface: identical to pi05_droid (exterior + wrist image at 224 px, joint and gripper position, prompt; 15 x 8
joint-velocity + gripper actions).